Success stories
Here are some clients who trust us for their digital transformation.
Logistics
We optimize operating costs through the selection of the best routes, times and movements.
Smart Fishing
We use AI models to monitor fish populations, predict their growth and even their migratory patterns.
LegalTech
We create software specialized in legal terminology. Beyond digitizing documents, we classify information optimally.
HR
Get to know your team's perceptions in depth to develop tailored action plans and optimize your results.
Banking
We leverage big data processing to segment customers, prevent fraud, generate insights and develop business strategies.
Health
Together with prestigious institutions in the healthcare sector, promoting research and the improvement of quality of life through the use of AI.
SmartAgriculture
Find out about our collaboration in the project to identify the best time to harvest the olive grove.
- Marketing
CTR PREDICTION
OBJECTIVES
In 2017, one of the world’s leading automobile manufacturers and the largest carmaker in Europe, needed to measure the return on investment (ROI) and predict the performance of data and campaigns to maximize advertising spend by predicting Click Through Ratio (CTR).
CHALLENGES
Design a very efficient model to exploit the large amount of data (150M impressions - 2 months of campaign), structured as categorical variables, with very unbalanced classes.
RESULTS
Creation of a personalized and predictive CTR model with adservers data, analyzing more than 100 million monthly interactions between users and ads with F1-score metrics: 70%.
Cognitive basis for recommending an ad in a precision marketing context.
SALES PREDICTION
OBJECTIVES
In 2018, the leading global foodservice retailer, with sales of more than €950 M in Spain, wanted to improve predictive marketing accuracy and specifically, measure return on investment (ROI) by predicting weekly sales of its franchises (>420), and maximize advertising campaigns by controlling investment by channel.
CHALLENGES
Large amounts of unstructured data from different sources (4 years of sales, advertising campaigns, weather, social and sporting events, vacations, etc.).
RESULTS
Decision support system that allows forecasting sales for the coming weeks with 96% accuracy, based on environmental variables and investment levels in different advertising channels.
Identification of the weight of each channel in sales (Simulation of counterfactuals) that maximizes the return of advertising campaigns, helping to decide the allocation by channels and scheduling over time.
- SMART FISHING
SMART ROUTE PREDICTION
OBJECTIVES
Satlink, a leading global provider of technology solutions for the fishing segment, wanted to add value to its marketed products by predicting, with a medium-term time horizon (7-10 days), the future trajectory of a drifting buoy and prevent it from colliding with the shore in order to increase its lifetime.
CHALLENGES
Incorporation of a large amount of daily oceanographic data (currents, wind, waves, temperature, etc.), which may fail or be delayed due to external events.
High uncertainty in weather forecasts.Robust service deployment, used daily by multiple customers.
RESULTS
Adding value to your marketed products with ocean information and increasing buoy life by 186%.
Information monitoring alert system that allows multiple daily attempts to try to recover from incidents, with multiple concurrent requests.
PREDICTION OF THE QUANTITY OF FISH AT THE BEACON
OBJECTIVES
Fishing prediction from echo sounder signals and oceanographic data from beacons.
Identify markings that allow discrimination between the most frequently caught tuna species.
CHALLENGES
Large amount of data to feed the models: beacon position information (latitude, longitude, ocean), echosounder (individual layer tags) and surface oceanography (~90 million tags from ~20k beacons over 4 years).
RESULTS
Fishing detection with 95% accuracy.
Development of classification models to predict the presence/absence of tuna under the beacon and regression models to predict the exact total amount of tons of tuna under the beacon.
ELECTRONIC MONITORING
OBJECTIVES
Satlink currently has an electronic monitoring system for fishing vessels with video cameras that record the relevant activity on deck. The recordings are processed with basic computer vision algorithms to detect the different fish species and facilitate the task of the experts when reviewing the large amount of hours of video.
CHALLENGES
Lack of labeled videos to solve the object tracking problem. This makes it complicated to train specific models that can take into account the temporal component.
RESULTS
Development and implementation of the most appropriate tracking algorithm, taking advantage of existing object detection algorithms. Reduction in the number of false positives by taking into account the temporal component. The developed algorithms also serve as assistants for tagging new videos, if necessary.
- BANKING
CUSTOMERS CLUSTERIZATION
OBJECTIVES
A bank with more than 3 million customers wanted to segment its customers based on their bank card transactions and according to five areas (how, where, when, when, how much, what) and to chart trends in the evolution of consumption patterns.
CHALLENGES
Large volume of data with over 3 million customers with a number of transactions semi-annually around, ~190M banking transactions, with ~40 columns, need for business sense cluster interpretation.
RESULTS
Scalable clustering models that serve as a cognitive foundation and input to the bank's other models: income prediction, propensity to consume, and customer and merchant similarity measures.
Build customer loyalty by offering more personalized products and services and selecting the right marketing channels for each audience.
CREDIT CARD FRAUD DETECTION
OBJECTIVES
Given a bank card transaction, develop a system to manage alerts and real-time prediction of whether it is fraudulent or legitimate.
Development of a decision support system to manage alerts. That is, deciding when to alert or suspend an operation taking into account all the costs involved.CHALLENGES
Large data set (+150 million payment transactions). Need to perform model inference in real time ( in order to avoid customer lag at checkout).
Need to perform model inference in real time (so that the customer does not perceive a delay when paying).RESULTS
36% improvement over previous accuracy.
Fraud classification models improving accuracy and keeping false positive rate (customer inconvenience) under control.
With the information gathered from the project data, a study was conducted on the behavior of the fraudsters.- HEALTH
STRUCTURING OF BREAST CANCER REPORTS
OBJECTIVES
Development, implementation and deployment of several Natural Language Processing (NLP) models aimed at structuring and coding clinical reports of breast cancer biopsies in SNOMED terminology, that would allow the Hospital Puerta del Mar de Cádiz to save time of its physicians and increase the productivity of the Anatomic Pathology service.
CHALLENGES
Unstructured language reports with multiple classes (400+).
RESULTS
Development of two independent coding models (based on SNOMED-CT terminology) and integration of the models with speech recognition.
CARDIOVASCULAR DISEASES STUDY
OBJECTIVES
The Instituto de Salud Carlos III together with Quirón Prevención, which provided the anonymous data used in this study and supported the analyses and interpretation, wished to determine the influence of different factors on the risk of cardiovascular disease.
CHALLENGES
Maintain the highest scientific standards to be accepted in prestigious journals; using non-standardized data from many sources; with more than 500,000 patients with multiple medical recognitions over the years.
RESULTS
Publication of study results in the European Journal of Preventive Cardiology. Through the homogenization of textual information, evidence is provided on how physical activities can reduce the risk of serious diseases.
- LEGALTECH
NLP IN END-TO-END LEGAL TEXTS
OBJECTIVES
In 2018, the leading company in Spain in the provision of legal information, wanted to classify, tag and identify key data automatically from a large corpus (1M+) of constantly growing judgments and legal documents and be able to manage the search needs on these documents for decision making.
CHALLENGES
Start with unlabeled data, which given the very specific legal and Spanish language context makes the use of pre-trained tools impossible. Requires use of annotation tools and active learning strategy and SOTA tools in 2018.
RESULTS
Classification system of judgments and documents according to legal categories organized by numerous hierarchical classification levels with an accuracy of more than 90%.
Semantic search is 22 times more efficient than the previous solution, which significantly reduces operational time and increases service productivity.
- HR
SENTIMENT ANALYSIS
OBJECTIVES
The client in the digital sector wanted to develop a tool that would allow them to collect texts from any source to analyze the information, understand it and take action on the work environment and monitor the opinion, emotion or attitude of employees.
CHALLENGES
Need for robust model with high accuracy in order to increase the % of correctly classified texts without adding computational cost.
RESULTS
Development of a multi-label text classification and sentiment analysis model with 96% accuracy that helps to adopt new strategies and make decisions in an agile way. Use of state of the art models in NLP (Transformers): contribution to Hugging Face. Active learning techniques: zero-shot learning, entropy-based intelligent labeling.
ROUTE OPTIMIZATION SYSTEM
OBJECTIVES
Analysis and optimization of the routing system to provide it with more flexibility, modularity and maintainability.
In addition, the previous system was to be upgraded by correcting design errors and improving performance.
CHALLENGES
Code not very maintainable and difficult to understand as it is not documented and absence of a global vision of the optimization problem to be solved from a mathematical point of view.
RESULTS
New formulation of the mathematical problem for route management that allows adapting to new scenarios and requirements as the product is implemented in each country.
Improvement of the route optimization system to minimize operational costs, planning time, fuel use and pollution, improving traceability and fleet management.
IDENTIFICATION OF THE OPTIMAL HARVEST TIME IN OLIVE GROVES
OBJECTIVES
Automatic and accurate identification of the optimal harvest time in olive groves in order to maximize the production of the highest quality olive oil, virgin and extra virgin.
CHALLENGES
Construction of a Smart-Agriculture tool based on Smart-Agriculture, massive data processing of the olive value chain. This tool relies on different technologies to obtain, store and process large amounts of data from various sources.
RESULTS
The future BeHTool platform builds three fundamental components:
- Obtaining information coming from sensor networks, satellites and weather stations.
- Management of all the above information and combination with data from olive groves and mills, such as phenological state and production.
- Advanced analytical capabilities and predictive models that, based on the data collected and stored in the previous modules, issue predictions.